The Intelligent Probe – Redefining Transesophageal Echocardiography in the Era of Artificial Intelligence

AI - definition and applications in TEE

AI is defined as the ability of computer systems to perform tasks that usually require human levels of intelligence.^[2] It refers to algorithms that allow computers to make decisions based on patterns uncovered from data. Machine learning (ML) is one of the subsets of AI which allows systems to analyze data, recognize patterns, and generate predictions without being explicitly told what to do. ML encompasses supervised and unsupervised learning. In supervised learning, the system is taught using a labeled dataset with input features and their corresponding output. Using this, the algorithm makes predictions on new, unseen data. Or to put it simply, the algorithm “learns from experience.” In unsupervised learning, the system finds patterns in unlabeled data without prior exposure to the input datasets. This is more useful in data analysis, stratification, and reduction than in prediction.^[3] There exists another type of learning called “reinforcement learning” where, in contrast to supervised and unsupervised learning, the system learns to make decisions through trial and error. Deep learning is a subset of ML that uses artificial neural networks with multiple layers to learn from large amounts of data.

The TEE assessment requires meticulous image acquisition and analysis. In the vast majority of cardiac surgeries and percutaneous interventions, decision-making relies heavily on the TEE findings. Incorporating AI in TEE would provide faster and perhaps more accurate information with minimal user involvement and inter-user variability. However, recent critiques of AI emphasize distinguishing real capability from hype and warn against overconfidence in algorithmic predictions without strong empirical validation.

Anatomical intelligence in ultrasound is a machine-learning system, primarily developed by Philips Healthcare that can automatically identify and quantify anatomical images during ultrasound examination.

AI in echocardiography was first introduced by Siemens Healthineers which was followed by the EPIQ software from Philips Medical Systems.

TEE view classification

One of the major impediments in the application of deep learning strategies to TEE is the complexity and unstructured nature of the data. Steffner et al. demonstrated a deep learning model that was successful in accurately classifying 8 standardized TEE views.^[4] This could prove to be the stepping stone toward the successful integration of AI in TEE. In the SIMULATOR Randomized control trial, Pezel et al. showed that novices to TEE greatly benefited from simulation-based learning.^[5]

Left ventricular ejection fraction (EF)

Left ventricular EF, the percentage of blood ejected by the ventricle in every heartbeat, serves as the preliminary measure of left ventricular systolic function. It is regarded as one of the most reliable prognostic markers of cardiovascular disease and plays a critical role in the diagnosis and management of various cardiac diseases. During cardiac surgery, measurement of left ventricle (LV) EF forms a fundamental part of TEE assessment. The consensus by the American Society of Echocardiography recommends the modified Simpson’s rule (biplane method of disc summation) as the standard for EF assessment^[6] which requires the manual tracing of the end systolic and end-diastolic borders of the LV.

Auto EF is a platform application developed by GE Healthcare, which allows automated measurement of EF. In TEE, it requires the users to acquire specific views and the software automatically traces the endocardial borders for the calculation of EF. In a recent study by Borde et al.,^[7] it was found that the EF measured by auto EF showed strong agreement with that measured manually. This feature could potentially help save time in the measurement of EF and decrease inter-user variability in EF calculation. 4D LV Analysis is a vendor-independent software from Tomtec (now part of Philips Ultrasound Workspace), available on both TEE and TTE which allows semi-automated measurement of LV function including LV Volumes, LV mass, EF, and Myocardial strain. While automated EF software by other vendors is available and has been validated on TTE, no study validating them on TEE could be identified.

Myocardial strain analysis

Strain analysis is an advanced method for the assessment of Left and Right Ventricular Systolic function. Myocardial strain represents myocardial deformation during the various phases of the cardiac cycle. Speckle tracking echocardiography (STE) measures myocardial strain and expresses it as a percentage change in myocardial length during the cardiac cycle. While traditional methods of ventricular function assessment still hold good, they are limited by their reliance on ventricular geometric assumptions and their reliability on loading conditions.^[8] Strain analysis eliminates those drawbacks. Although 2D STE was initially used as a tool for LV strain analysis, its role has now been expanded to measure LA and RV strain also.

Automated functional imaging, an application from GE Healthcare, allows for myocardial strain assessment by 2D-STE.^[9] AutoStrain by Tomtec is an AI-powered software for strain analysis that utilizes two automation technologies: Auto View recognition and Auto Contour placement with Speckle tracking for rapid strain analysis. This is a vendor-independent software. Strain analysis of LV, RV, and LA can be performed with this software. The Auto View Recognition identifies the TEE view from the acquired cine loop and labels it. The Auto Contour placement automatically traces the endocardial border. The user can, if required, revise the borders. Strain values are then measured by the software. A study by Peng et al. found good agreement between the values of LV strain measured manually and those measured by the automatic software on TTE.^[10]

Auto mitral annular plane systolic excursion (MAPSE)

MAPSE is a parameter used to assess LV systolic function. It measures LV longitudinal shortening, which has been established as a sensitive parameter that reflects the LV systolic function. It has a high correlation with global longitudinal strain and is suggested as a surrogate for LVEF. Berg et al. developed a continuous neural network for automatically detecting the mitral valve and measuring MAPSE.^[11,12] Auto MAPSE comprises a convolutional neural network trained under supervised learning to detect the mitral valve annulus in TEE images. After detection of the mitral annulus, auto MAPSE calculates the distance traveled by the mitral valve from the highest to the lowest position in each wall of the LV and for every heartbeat. The software was validated and they further demonstrated that continuous monitoring of LV systolic function in a critical care setting was not only feasible with AutoMAPSE but also more precise than manual measurements.^[13] The application of this technology in real-time monitoring of LV systolic function in various clinical settings could assist clinicians in the management of critically ill patients.

Automated RV function assessment

Historically, RV has always received less attention than the left, until recently, when studies have shown overwhelming evidence that the right ventricle plays as significant a role in maintenance of hemodynamics as the left, though the LV does appear to bear the brunt of most cardiac diseases. The complex geometry of the right ventricle makes analysis of the right ventricular function more technically challenging than that of the LV. 3D echocardiography has been shown to provide good quantification of right ventricular volume and function, comparable to the information provided by cardiac magnetic resonance imaging. An automated algorithm that assessed RV volume and function using 3D TEE was found to be in good agreement with manual readings in a study by Nillesen et al.^[14]

4D auto RVQ by GE Healthcare is a fully automated tool that computes RV volumes and RV EF from acquired 4D datasets. In a study by Wu et al., it was found that the measurements provided by the software in TTE echo were in agreement with the values measured by Right Heart Catheterization, but only for high-quality images.^[15] Automated assessment of the right ventricle is challenging because of its complex anatomic characteristics, and consequently, literature on AI-based RV analysis is scarce.

Aortic valve analysis

Transcatheter aortic valve replacement (TAVR) has emerged as a popular alternative to surgical aortic valve replacement in high-risk patients. Patient selection for TAVR is a meticulous process that requires input from the Heart Team of the institution. Assessment of aortic dimensions and aortic valve geometry is one of the crucial steps in pre-procedure planning. Accurate data are usually obtained by computed tomography (CT) imaging. Development of semi-automated and automated software for aortic valve analysis on TEE has enabled accurate assessment of the valve in patients who are unable to undergo a CT scan.

eSie Valve from Siemens Healthcare is a software package which allows automated detection, quantification, and modeling of mitral and aortic valve anatomy. It uses the speckle tracking feature to detect the major landmarks associated with each valve. It generates the aortic valve annulus by area or by perimeter. The user can measure the aortic dimensions at either a single point in the cardiac cycle or across the entire cardiac cycle.^[16] The measured values were found to correlate well with the values measured on CT.^[16,17]

Aortic Valve Navigator by Philips Healthcare is another automated software that enables assessment of the aortic valve dimensions from the acquired 3D TEE images. It has been found to be reproducible with very low inter-observer variability and relatively fast in providing the measurements. However, its performance remains dependent on the quality of the 3D images acquired. The measurements provided were found to correspond to the CT measurements.^[18]

4D auto AVQ, an automated software from GE Healthcare, was found to reliably estimate the size of the aortic annulus in patients undergoing TAVR in a cohort study by Massie et al. The automated software provided results that were accurate and reproducible and required minimal user effort.^[19]

Mitral valve assessment

The importance of echocardiography in the diagnosis and therapeutic management of patients with mitral valve disease cannot be emphasized enough. A comprehensive assessment of the valve anatomy and its quantitative parameters is essential before management of the patient, particularly considering the growing popularity of percutaneous interventions, where direct visualization of the valve would not be possible, unlike during surgery. From manual assessment to semi-automated methods that nevertheless required considerable input from the user, we now have fully automated software that provides a complete assessment of the valve with minimal effort put in by the user. Jeganathan et al. reviewed the automated eSie Valve Software (from Siemens), which required the acquisition of 3D cine loops of the mitral valve from which the algorithm automatically detected and traced the mitral valve. It then provided the measurements, including (i) mitral annulus anterolateral posteromedial diameter, (ii) mitral annulus anteroposterior diameter, (iii) mitral annular area, (iv) mitral annulus non-planarity angle, (v) mitral annulus total perimeter, and (vi) anterior and posterior leaflet areas. The authors concluded that the automated measurements were accurate with good reproducibility.^[20] 3D Auto MV by Tomtec is another fully automated software that provides all the requisite mitral valve measurements with minimal user effort. Caution should, however, be exercised because skepticism still prevails regarding the accuracy of the measurements provided by fully automated software, regardless of the vendor and user discernment is of the essence.

Tricuspid valve analysis

Most automated valve analysis software is restricted to the analysis of only the mitral and the aortic valves. With deepening understanding of Functional Tricuspid Regurgitation and its impact on patient outcomes, there has been a renewed interest in developing software for the assessment of the tricuspid valve.

Autovalve Analysis by Siemens Healthcare is a semiautomated, vendor-independent software that can assess the mitral, aortic, and tricuspid valves. It allows a fully automated identification of various structures and provides static and dynamic analysis of the valves. It was also found to provide simultaneous assessment of the mitral and aortic valves alongside the tricuspid valve, allowing evaluation of the spatial and geometric configuration of all the valves through the cardiac cycle. The software was found to be reproducible and accurate with minimal user effort required.^[21]

AutoTV by Philips Healthcare is a tricuspid valve annulus quantification tool with 14 automated 3D annulus measurements and 10 dedicated measurements for percutaneous device procedure planning and size re-confirmation.^[22] 4D Auto TVQ (GE Healthcare) is a semi-automated software that allows quantification of the Tricuspid Valve on both TTE and TEE.^[23]

3D color flow quantification

Assessment of the severity of mitral regurgitation (MR) by echocardiography requires an integrated qualitative, semi-quantitative, and quantitative approach. The process can get quite cumbersome, and the different 2D methods of quantifying MR come with their own limitations. Auto color flow quantification (Auto CFQ) by Philips Healthcare provides an automated quantification of mitral valve regurgitant volume (RVol) based on 3D datasets that have been acquired. According to the vendor validation study cited in the Philips white paper, the automated mitral RVol correlated well with the measurements obtained using cardiac magnetic resonance. The measurements provided include RVol (ml) and peak flow rate (mL/s) along with a graphical illustration of the regurgitant flow over time. At present, the only validation study available is the one performed by the manufacturer.^[24]

Fusion imaging

With transcatheter interventions rapidly gaining popularity in the management of cardiovascular conditions, software that combines transesophageal echocardiographic datasets with live fluroscopic images has been developed. This allows the interventionist to be able to visualize the landmarks on the fluoroscopic field, improving navigation and precision while enabling a reduction in radiation exposure.^[25] It requires that the echocardiographic and fluoroscopic equipment should be from the same vendor for compatibility. EchoNavigator by Philips and Syngo TrueFusion by Siemens are the commercially available software for fusion imaging.

As the integration of AI with TEE is still in its nascent stages, there is a paucity of peer-reviewed literature validating the clinical accuracy and utility of the vendor-provided software. Tables 1-4 summarize automated software provided by different vendors along with a brief account of the validation study where applicable.^[26-37] Availability of AI tools may vary depending on the ultrasound platform, software version, and probe compatibility [Figures 1 and 2].^[38-53]

Close

Figure 1:

Integration of artificial intelligence into perioperative transesophageal echocardiography workflow. OT: Operation theater, TEE: Transesophageal echocardiography.

Export to PPT

Close

Figure 2:

Levels of evidence supporting artificial intelligence applications in transesophageal echocardiography.

Export to PPT